pith. machine review for the scientific record. sign in

arxiv: 2605.00675 · v1 · submitted 2026-05-01 · 💻 cs.CV

Recognition: unknown

DMDSC: A Dynamic-Margin Deep Simplex Classifier for Open-Set Recognition on Medical Image Datasets

Vishal , Arnav Aditya , Nitin Kumar , Saurabh J. Shigwan
Authors on Pith no claims yet

Pith reviewed 2026-05-09 19:43 UTC · model grok-4.3

classification 💻 cs.CV
keywords open-set recognitionmedical imagingclass imbalancedeep simplex classifierdynamic marginneural collapse
0
0 comments X

The pith

Dynamic margins based on label frequency improve open-set recognition for rare medical conditions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Medical imaging datasets suffer from severe class imbalance, with rare pathologies appearing far less often than common ones. The paper proposes DMDSC, which modifies the deep simplex classifier by making the margin between classes dynamic and dependent on how frequently each class appears in the data. This change imposes stronger separation requirements on rare classes to produce tighter feature clusters. The result is better accuracy on known classes and more reliable rejection of unknown samples compared to methods that use the same margin for all classes.

Core claim

The DMDSC framework features a dynamic margin approach that automatically adapts class-specific margins based on label frequency. This enforces a higher penalty and tighter feature clustering for rare pathologies to counteract the effects of data imbalance in open-set recognition on medical image datasets. Experiments on BloodMNIST, OCTMNIST, DermaMNIST, and BreaKHis show it outperforms state-of-the-art methods.

What carries the argument

The dynamic margin scaled according to label frequency within the Deep Simplex Classifier structure, which uses Neural Collapse to maximize inter-class separation but now with class-dependent margins.

If this is right

  • The model achieves superior performance on open-set recognition tasks across multiple medical imaging benchmarks with class imbalance.
  • Rare pathologies receive tighter feature clustering due to increased margin penalties.
  • Known class classification accuracy is maintained while improving rejection of unknown samples.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This approach of frequency-dependent margins could be tested in non-medical domains with similar imbalance issues.
  • Integrating it with uncertainty-aware variants of the simplex classifier might yield additional gains in clinical applications.

Load-bearing premise

Scaling the margin by label frequency alone will produce tighter and more separable clusters for rare classes without introducing overfitting to the frequency statistics or reducing performance on common classes.

What would settle it

Training the DMDSC model and a fixed-margin baseline on a medical dataset like BloodMNIST, then checking if rare class features show reduced variance and if unknown sample rejection rates increase without harming common class accuracy.

Figures

Figures reproduced from arXiv: 2605.00675 by Arnav Aditya, Nitin Kumar, Saurabh J. Shigwan, Vishal.

Figure 1
Figure 1. Figure 1: Illustration of the proposed open-set recognition framework with dynamic mar￾gin learning. In our proposed framework, input images are mapped to a hyperspherical embedding space where class centers (stars) are fixed at the vertices of a simplex ETF. The distance between any two class centers is d, ensuring maximal inter-class separa￾tion. Known-class samples (dots) are pulled toward their corresponding cen… view at source ↗
Figure 2
Figure 2. Figure 2: Sample images from the five datasets (a) BloodMNIST (b) OCTMNIST (c) BreaKHis (d) DermaMNIST (e) Augmented Skin Conditions 4.2 Evaluation Metrics We use accuracy (ACC) to evaluate closed-set classification performance. To measure open-set performance, we report AUROC, which is a threshold-independent [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation studies on margin parameters. Each result is the average of five runs [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
read the original abstract

Medical imaging datasets are often characterized by extreme class imbalances, where rare pathologies are significantly underrepresented compared to common conditions. This imbalance poses a dual challenge for Open-Set Recognition (OSR): models must maintain high classification accuracy on known classes while reliably rejecting unknown samples unseen during training in the clinical settings. While recently proposed Deep Simplex Classifier (DSC)~\cite{cevikalp2024reaching} and UnCertainty-aware Deep Simplex Classifier (UCDSC)~\cite{Aditya_2026_WACV} successfully leverage Neural Collapse to ensure maximal inter-class separation, they rely on a uniform margin that does not account for the varying densities of medical classes. In this paper, we propose DMDSC an enhanced framework featuring a dynamic margin approach. Our approach automatically adapts class-specific margins based on label frequency, enforcing a higher penalty and tighter feature clustering for rare pathologies to counteract the effects of data imbalance. Extensive experiments conducted on diverse medical benchmarks on BloodMNIST\cite{medmnistv2}, OCTMNIST\cite{medmnistv2}, DermaMNIST\cite{medmnistv2}, and BreaKHis~\cite{spanhol2015dataset} datasets, demonstrate that our framework outperforms state-of-the-art methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes DMDSC, an extension of the Deep Simplex Classifier (DSC) framework that introduces class-specific dynamic margins scaled by label frequency to better handle extreme class imbalance in open-set recognition for medical images. It claims this enforces tighter feature clustering for rare pathologies and yields superior performance over prior DSC/UCDSC variants and other SOTA methods on the BloodMNIST, OCTMNIST, DermaMNIST, and BreaKHis benchmarks.

Significance. If the empirical gains hold under rigorous validation, the dynamic-margin idea offers a lightweight, frequency-aware modification to Neural-Collapse-based classifiers that could improve rare-class detection and unknown rejection in clinically imbalanced settings without requiring architectural changes.

major comments (2)
  1. [Method] The central justification—that scaling the simplex margin inversely with label frequency produces measurably tighter within-class collapse for rare classes while preserving the Neural-Collapse fixed point—lacks any derivation, fixed-point analysis, or gradient-magnitude study. No section shows that the non-uniform margin schedule keeps the attractive fixed point intact or avoids destabilizing optimization for high-frequency classes.
  2. [Experiments] The abstract asserts outperformance on four datasets, yet the provided text supplies neither quantitative tables, ablation results isolating the dynamic-margin component, nor statistical tests. Without these, the load-bearing empirical claim cannot be assessed.
minor comments (1)
  1. [Abstract] The abstract states 'extensive experiments' but reports no key metrics, which reduces its utility as a standalone summary.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. The comments highlight important areas for strengthening the theoretical justification and empirical presentation. We address each major comment point by point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Method] The central justification—that scaling the simplex margin inversely with label frequency produces measurably tighter within-class collapse for rare classes while preserving the Neural-Collapse fixed point—lacks any derivation, fixed-point analysis, or gradient-magnitude study. No section shows that the non-uniform margin schedule keeps the attractive fixed point intact or avoids destabilizing optimization for high-frequency classes.

    Authors: We agree that the submitted manuscript lacks an explicit derivation and fixed-point analysis for the dynamic-margin schedule. In the revision we will add a dedicated subsection deriving the modified Neural Collapse objective under class-specific margins. The analysis will show that the frequency-scaled margin preserves the attractive fixed point by scaling the within-class variance term proportionally to inverse frequency, while the inter-class separation term remains unchanged. We will also include a gradient-magnitude study demonstrating that the schedule does not destabilize optimization for high-frequency classes, as their gradients remain bounded by the original uniform-margin case. revision: yes

  2. Referee: [Experiments] The abstract asserts outperformance on four datasets, yet the provided text supplies neither quantitative tables, ablation results isolating the dynamic-margin component, nor statistical tests. Without these, the load-bearing empirical claim cannot be assessed.

    Authors: We acknowledge that the quantitative tables, ablation studies isolating the dynamic margin, and statistical tests were insufficiently detailed or formatted in the submitted version. The revision will prominently include full performance tables for BloodMNIST, OCTMNIST, DermaMNIST, and BreaKHis, with direct comparisons to DSC, UCDSC, and other SOTA baselines. We will add a dedicated ablation subsection varying only the margin schedule (uniform vs. dynamic) while keeping all other components fixed, and report paired statistical tests (e.g., McNemar or t-tests with p-values) across multiple runs to support the outperformance claims. revision: yes

Circularity Check

0 steps flagged

Dynamic margin extension on prior DSC relies on self-citation for Neural Collapse but adds independent frequency-based rule

full rationale

The manuscript builds DMDSC directly on DSC (cevikalp2024reaching) and UCDSC (Aditya_2026_WACV) to leverage Neural Collapse for simplex separation, then introduces a new dynamic margin scaled by label frequency. One citation (UCDSC) has author overlap, qualifying as minor self-citation, but it is not load-bearing for the central novelty: the frequency-driven margin adaptation itself is presented as an explicit design choice without any equation that reduces the claimed tighter clustering or OSR gain to a fitted parameter or self-defined quantity. No derivation chain collapses by construction; the paper remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit equations or derivations; the ledger is therefore empty. The implicit reliance on Neural Collapse geometry from the cited prior work is noted but not formalized here.

pith-pipeline@v0.9.0 · 5540 in / 1094 out tokens · 53139 ms · 2026-05-09T19:43:28.709654+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 6 canonical work pages

  1. [1]

    Mendeley Data, V1 (2020).https://doi.org/10

    Acevedo, A., Merino, A., Alférez, S., Cabrera, J.R., Pereira, C., León, A., Sánchez, P.: A dataset for microscopic peripheral blood cell images for development of au- tomatic recognition systems. Mendeley Data, V1 (2020).https://doi.org/10. 17632/snkd93bnjr.1,https://doi.org/10.17632/snkd93bnjr.1

    work page doi:10.17632/snkd93bnjr.1 2020

  2. [2]

    Data in Brief30, 105474 (2020).https://doi.org/ 10.1016/j.dib.2020.105474

    Acevedo, A., Merino, A., Alférez, S., Cabrera, J.R., Pereira, C., León, A., Sánchez, P.: A dataset of microscopic peripheral blood cell images for development of au- tomatic recognition systems. Data in Brief30, 105474 (2020).https://doi.org/ 10.1016/j.dib.2020.105474

    work page doi:10.1016/j.dib.2020.105474 2020

  3. [3]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Aditya, A., Kumar, N., Shigwan, S.: UCDSC: Open set uncertainty aware deep simplex classifier for medical image datasets. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 4787–4796 (March 2026)

    2026

  4. [4]

    Bendale, A., Boult, T.E.: Towards open world recognition. CVPR pp. 1893–1902 (2015)

    1902

  5. [5]

    IEEE Transactions on Neural Networks and Learning Systems36(5), 8178–8191 (2024)

    Cevikalp, H., Saribas, H., Uzun, B.: Reaching nirvana: Maximizing the margin in both euclidean and angular spaces for deep neural network classification. IEEE Transactions on Neural Networks and Learning Systems36(5), 8178–8191 (2024)

    2024

  6. [6]

    Pattern Recognition138, 109385 (2023)

    Cevikalp, H., Uzun, B., Salk, Y., Saribas, H., Köpüklü, O.: From anomaly detection to open set recognition: Bridging the gap. Pattern Recognition138, 109385 (2023)

    2023

  7. [7]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(11), 8065–8081 (2021)

    Chen, G., Peng, P., Wang, X., Tian, Y.: Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence44(11), 8065–8081 (2021)

    2021

  8. [8]

    In: European conference on computer vision

    Chen, G., Qiao, L., Shi, Y., Peng, P., Li, J., Huang, T., Pu, S., Tian, Y.: Learning open set network with discriminative reciprocal points. In: European conference on computer vision. pp. 507–522. Springer (2020)

    2020

  9. [9]

    Chin, T.J., Cai, Z., Neumann, F.: Robust fitting in computer vision: Easy or hard? In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 701–716 (2018)

    2018

  10. [10]

    Skin Lesion Analysis Toward Melanoma Detection 2018: A Challenge Hosted by the International Skin Imaging Collaboration (ISIC)

    Codella, N., Rotemberg, V., Tschandl, P., Celebi, M.E., Dusza, S., Gutman, D., Helba,B.,Kalloo,A.,Liopyris,K.,Marchetti,M.,etal.:Skinlesionanalysistoward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368 (2019)

    work page Pith review arXiv 2018

  11. [11]

    Ad- vances in Neural Information Processing Systems31(2018)

    Dhamija, A.R., Günther, M., Boult, T.: Reducing network agnostophobia. Ad- vances in Neural Information Processing Systems31(2018)

    2018

  12. [12]

    In: British Machine Vision Conference

    Ge, Z., Demyanov, S., Chen, Z., Garnavi, R.: Generative openmax for multi-class open set classification. In: British Machine Vision Conference. BMVA Press (2017)

    2017

  13. [13]

    IEEE TPAMI43(10), 3614–3631 (2021)

    Geng, C., Huang, S.J., Chen, S.: Recent advances in open set recognition: A survey. IEEE TPAMI43(10), 3614–3631 (2021)

    2021

  14. [14]

    IEEE transactions on pattern analysis and machine intelligence43(10), 3614–3631 (2020)

    Geng, C., Huang, S.j., Chen, S.: Recent advances in open set recognition: A survey. IEEE transactions on pattern analysis and machine intelligence43(10), 3614–3631 (2020)

    2020

  15. [15]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016)

    2016

  16. [16]

    Deep Anomaly Detection with Outlier Exposure

    Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)

    work page Pith review arXiv 2018

  17. [17]

    In: Proceedings of the 7th International Conference on Learning Repre- sentations (ICLR) (2019),https://openreview.net/forum?id=HyxCxhRcY7 16 Vishal et al

    Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. In: Proceedings of the 7th International Conference on Learning Repre- sentations (ICLR) (2019),https://openreview.net/forum?id=HyxCxhRcY7 16 Vishal et al

    2019

  18. [18]

    Intelligent data analysis6(5), 429–449 (2002)

    Japkowicz, N., Stephen, S.: The class imbalance problem: A systematic study. Intelligent data analysis6(5), 429–449 (2002)

    2002

  19. [19]

    Cell172(5), 1122– 1131.e9 (2018).https://doi.org/10.1016/j.cell.2018.02.010

    Kermany, D.S., Goldbaum, M., Cai, W., Valentim, C.C., Liang, H., Baxter, S.L., McKeown, A., Yang, G., Wu, X., Yan, F., Dong, J., et al.: Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell172(5), 1122– 1131.e9 (2018).https://doi.org/10.1016/j.cell.2018.02.010

    work page doi:10.1016/j.cell.2018.02.010 2018

  20. [20]

    Scientific Reports15(1), 22617 (2025)

    Lin, Y., He, S., Luo, W.: Dynamic margin contrastive learning for open-set recog- nition in long-tailed sonar imagery. Scientific Reports15(1), 22617 (2025)

    2025

  21. [21]

    In: International conference on medical image computing and computer-assisted intervention

    Liu, M., Xu, L., Zhang, J.: Learning large margin sparse embeddings for open set medical diagnosis. In: International conference on medical image computing and computer-assisted intervention. pp. 548–558. Springer (2023)

    2023

  22. [22]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Miller, D., Sunderhauf, N., Milford, M., Dayoub, F.: Class anchor clustering: A loss for distance-based open set recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3570–3578 (2021)

    2021

  23. [23]

    In: European conference on computer vision

    Moon, W., Park, J., Seong, H.S., Cho, C.H., Heo, J.P.: Difficulty-aware simulator for open set recognition. In: European conference on computer vision. pp. 365–381. Springer (2022)

    2022

  24. [24]

    Kaggle (2023),https:// www.kaggle.com/datasets/syedalinaqvi/augmented-skin-conditions-image- dataset

    Naqvi, S.A.R.: Augmented skin conditions image dataset. Kaggle (2023),https:// www.kaggle.com/datasets/syedalinaqvi/augmented-skin-conditions-image- dataset

    2023

  25. [25]

    In: ECCV

    Neal, L., et al.: Open set recognition with counterfactual images. In: ECCV. pp. 613–628 (2018)

    2018

  26. [26]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Oza, P., Patel, V.M.: C2ae: Class conditioned auto-encoder for open-set recogni- tion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2307–2316 (2019)

    2019

  27. [27]

    Proceedings of the National Academy of Sciences117(40), 24652–24663 (2020)

    Papyan, V., Han, X., Donoho, D.L.: Prevalence of neural collapse during the ter- minal phase of deep learning training. Proceedings of the National Academy of Sciences117(40), 24652–24663 (2020)

    2020

  28. [28]

    IEEE Transactions on Pattern Analysis and Machine Intelligence40(3), 762–768 (2018)

    Rudd, E.M., et al.: The extreme value machine. IEEE Transactions on Pattern Analysis and Machine Intelligence40(3), 762–768 (2018)

    2018

  29. [29]

    IEEE TPAMI36(11), 2317–2324 (2014)

    Scheirer, W.J., Jain, L.P., Boult, T.E.: Probability models for open set recognition. IEEE TPAMI36(11), 2317–2324 (2014)

    2014

  30. [30]

    IEEE transactions on pattern analysis and machine intelligence35(7), 1757–1772 (2012)

    Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence35(7), 1757–1772 (2012)

    2012

  31. [31]

    Ieee transactions on biomedical engineering 63(7), 1455–1462 (2015)

    Spanhol, F.A., Oliveira, L.S., Petitjean, C., Heutte, L.: A dataset for breast cancer histopathological image classification. Ieee transactions on biomedical engineering 63(7), 1455–1462 (2015)

    2015

  32. [32]

    IEEE transactions on pattern analysis and machine intelligence30(11), 1958–1970 (2008)

    Torralba, A., Fergus, R., Freeman, W.T.: 80 million tiny images: A large data set for nonparametric object and scene recognition. IEEE transactions on pattern analysis and machine intelligence30(11), 1958–1970 (2008)

    1958

  33. [33]

    Scientific data5(1), 180161 (2018)

    Tschandl, P., Rosendahl, C., Kittler, H.: The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data5(1), 180161 (2018)

    2018

  34. [34]

    Pattern Recogni- tion126, 108564 (2022)

    Wang, Z., Dong, Q., Guo, W., Li, D., Zhang, J., Du, W.: Geometric imbalanced deep learning with feature scaling and boundary sample mining. Pattern Recogni- tion126, 108564 (2022)

    2022

  35. [35]

    IEEE Access12, 122852–122877 (2024).https://doi.org/10.1109/ACCESS.2024.3442569 DMDSC 17

    Xu, Y., Wang, R., Zhao, R.W., Xiao, X., Feng, R.: Semi-supervised and class- imbalanced open set medical image recognition. IEEE Access12, 122852–122877 (2024).https://doi.org/10.1109/ACCESS.2024.3442569 DMDSC 17

    work page doi:10.1109/access.2024.3442569 2024

  36. [36]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Yang, H.M., Zhang, X.Y., Yin, F., Liu, C.L.: Robust classification with convolu- tional prototype learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3474–3482 (2018)

    2018

  37. [37]

    International Journal of Computer Vision132(12), 5635–5662 (2024)

    Yang, J., Zhou, K., Li, Y., Liu, Z.: Generalized out-of-distribution detection: A survey. International Journal of Computer Vision132(12), 5635–5662 (2024)

    2024

  38. [38]

    Scientific Data8(1), 1–14 (2021)

    Yang, J., Shi, Y., Ni, B., et al.: Medmnist v2: A large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data8(1), 1–14 (2021)

    2021

  39. [39]

    In: CVPR

    Yoshihashi, R., et al.: Classification-reconstruction learning for open-set recogni- tion. In: CVPR. pp. 4016–4025 (2019)

    2019